Setting and ethics
This retrospective, multicenter study was performed in three intensive care units (ICUs) at two university medical centers (Unité de Réanimation Cardio-Vasculaire, Institut Cœur-Poumon, Lille University Medical Center, Lille, France; Pôle des Réanimations, Hôpital Salengro, Lille University Medical Center, Lille, France; Unité de Réanimation Cardio-Thoracique, Vasculaire et Respiratoire, Amiens University Medical Center, Amiens, France). The study was approved by the investigational review board at the French Society of Anesthesia, Intensive Care and Perioperative Medicine (Paris, France) on January 17, 2022 (reference: CERAR IRB 00,010,254-2022-009). In accordance with French legislation, all datasets used in the present study were registered with the French National Data Protection Commission (Commission nationale de l'informatique et des libertés (Paris, France; methodology MR-004; reference: 2208336v0 for Amiens University Medical Center, and DEC2015-14 for Lille University Medical Center) [16].
Study population
Consecutive adult patients on VA-ECMO support for refractory CS rated as Interagency Registry for Mechanically Assisted Circulatory Support profile 1 or 2 (“crash and burn” and “progressive decline on inotropic support,” respectively) or Society for Cardiovascular Angiography and Interventions stage D or E (“deteriorating” and “extremis,” respectfully) between January 2013 until January 2020 were considered for inclusion [17, 18].
ECMO initiated exclusively for high-risk percutaneous coronary intervention, cardiopulmonary resuscitation, or an intraoperative procedure was not included in the study. The formal exclusion criteria were age under 18, a moribund patient (death within 48 h of initiating VA-ECMO), missing data for PaO2, uncertainty as to whether a blood sample came from the right arterial or ulnar arteries (in cases of femoral artery cannulation), a second ECMO run in the same patient, central cannulation, and right ventricle to pulmonary artery ECMO.
Data collection
Clinical data and outcome data were gathered from paper-based or electronic medical records [Sillage (SIB, Rennes, France) and IntelliSpace Critical Care and Anesthesia (Philips Healthcare, Koninklijke Philips N.V., the Netherlands) for the Lille centers and Centricity Critical Care (formerly known as Clinisoft) software (GE Healthcare, Barrington, IL) for the Amiens center. Laboratory information was collected from devoted software applications [Molis® (CompuGroup Medical, Koblenz, Germany) in the Lille centers and Clinisoft (GE Healthcare, Barrington, IL) in the Amiens center]. We collected anthropometric data, a detailed medical history, the Simplified Acute Physiologic Score (SAPS II), the baseline arterial blood lactate level, and other laboratory variables. We also collected all relevant information concerning ECMO management, including the indication for ECMO support, the cannulation site, the duration of cannulation, the type of device, the ECMO flow, and the occurrence of any ECMO-related complications.
Blood samples and laboratory analysis
Arterial blood samples were drawn into a 3-mL preheparinized syringe (BD Preset™, Plymouth, UK) from arterial lines positioned in the right radial artery. The samples were processed within minutes of collection on a point-of-care analyzer or sent to a central laboratory using an automatized pneumatic tube transportation system that shortened the laboratory delivery time to a few minutes. PaO2 was measured using an ABL90 FLEX or ABL800 FLEX blood gas analyzer (Radiometer® Medical ApS, Brønshøj, Denmark) or a GEM®Premier 4000 blood gas analyzer (Instrumentation Laboratory, Werfen, Bedford, MA, USA) depending on the center.
Management of VA-ECMO
The femoral artery or the right subclavian artery was cannulated percutaneously or with a semi-Seldinger approach. The femoral vein was cannulated by trained cardiovascular or thoracic surgeons. During the study period, three ECMO systems were in use: a Maquet ECMO system, comprising a Rotaflow centrifugal pump-based system and a Cardiohelp with a disposable 5.0/7.0HLS Set Advanced (Getinge AB, Göteborg, Sweden); a LivaNova system (LivaNova, Saluggia, Italy) with a revolution pump head; and a Eurosets system (Eurosets Srl, Medolla, Italy)].
Anticoagulation was initiated with a 100 IU/kg bolus of unfractionated heparin before cannulation for patients not on cardiopulmonary bypass prior to ECMO support, followed by a continuous infusion. The target was an anti-FXa level of 0.2–0.4 IU ml−1 in the Lille and Amiens centers.
The pump flow was adjusted to target a mean arterial pressure > 60 mmHg, SvO2 > 65% or ScVO2 > 70%, and aortic valve opening. Weaning was considered for recovery when the cardiac output was acceptable after reducing the ECMO flow and inotropic support to the minimum level.
It should be noted that PaO2 was adjusted at the discretion of the attending physician—primarily by changing the ECMO system’s fraction of inspired oxygen (FiO2) via an oxygen–air blender (Sechrist Industries, Anaheim, CA). The FiO2 on the ventilator was set to the minimum value and was modified as a function of the arterial oxygen saturation (SaO2) measured in the right upper limb arteries (for femoral cannulation) or the left upper limb arteries (for right subclavian arterial cannulation) or the cerebral near-infrared spectrometry index values when the SaO2 or peripheral oxygen saturation (SpO2) values were unavailable.
Definition of oxygen parameters
The following PaO2 values (in mmHg) were recorded or calculated over the first 48 h following admission to the ICU: (i) the three through PaO2 values on admission (Day 0), Day 1, and Day 2; (ii) the three peak PaO2 values on admission, Day 1, and Day 2; (iii) the mean daily peak PaO2 (mmHg), calculated as the mean of the daily peak values on admission, Day 1, and Day 2; (iv) the mean daily through PaO2, calculated as the mean of the daily through values; (v) the absolute peak PaO2, i.e., the highest peak PaO2 value between admission and Day 2; (vi) the overall mean PaO2, i.e., the mean of all PaO2 values between admission and Day 2; and (vi) the severity of hyperoxia, graded with reference to the overall mean PaO2 (mild: < 200 mmHg; moderate: 200–299 mmHg; severe: ≥ 300 mmHg).
Study endpoint
The primary study endpoint was 28-day all-cause mortality, as determined from the patient’s electronic medical records and the French national death registry (Institut national des statistiques et des études économiques, Paris, France) [19]. None of the study participants was lost to follow-up, and status at 28 days could be documented in all cases.
Statistical analysis
Data were presented as the mean ± standard deviation or the frequency (percentage), as appropriate. Non-survivors were compared with 28-day survivors using Student’s t test, a chi-squared test, or Fisher’s exact test, as appropriate. To assess the effects of ECMO flow rate and hemoglobin on PaO2, we used a multiple linear regression. Binary logistic regression analyses with 28-day mortality as the dependent variable were used to estimate the respective univariate associations with the absolute peak PaO2, mean daily peak PaO2, the overall mean PaO2, and the hyperoxia range. Unadjusted and adjusted odds ratios (ORs) with their 95% confidence interval (CI) were estimated from the binary logistic regression for a 10-point increment in PaO2-derived parameters. The results were adjusted for age, hypertension, the indication for VA-ECMO support, the SAPS II, and the arterial blood lactate on admission. Kaplan–Meier estimates were used for time-to-event analyses. Sensitivity analyses were performed by excluding patients with femoro-axillary ECMO and by using the OR as a measure of the effect size.
Missing data have been imputed for the following stages using predictive mean matching imputation (pmm). Five imputation data sets were produced. Missing data have been analyzed as missing-not-at-random according to the Rubin rule [20]. Prognostic variables related to 28-day mortality at the 20% in univariate analysis were included in the propensity score (regardless of their differences between the two groups): age, gender, hypertension, eGFR, diabetes, coronary disease, SAPS II, arterial lactate on admission, and etiology of refractory cardiogenic shock. For each patient, the probability to show hyperoxia has been estimated using logistic regression. For ATE (average treatment effect on the entire population) analysis, weights have been attributed to each patient of the hyperoxia (overall mean PaO2 > 150 mmHg) and no hyperoxia (overall mean PaO2 ≤ 150 mmHg) groups, making the two groups similar for the variables in the propensity score [21]. These weights were calculated using the stabilized inverse probability of treatment weighting (SIPTW). Balance (standardized mean differences lower than 10%) was checked for each variable. ORs were estimated before and after weighting on the propensity score using logistic regression. All tests were two-sided, and the threshold for statistical significance was set to p < 0.05. Statistical analysis was performed with R studio software for macOS (version 2021.09.1 + 372) and its «dplyr», «ggplot2», «survminer», «survival», « hrbrthemes», «tableone», « ggeffects», « WeightIt», « cobalt», « compareGroups», « mice» and « epiR» and “reshape2” packages.